fix: use session-pinned agent in RunStream instead of shared currentAgent

[dgageot]

Problem

Sub-agents dispatched via run_background_agent executed with root's toolset, model, and instructions. This happened because RunStream resolved the agent through the shared currentAgent field, which is never updated for background sessions.

Root Cause

Multiple places in the streaming loop (RunStream, processToolCalls, finalizeEventChannel, Summarize, and helper methods) independently called r.CurrentAgent() or r.CurrentAgentName(), which always returns the root agent for background sessions since run_background_agent never calls setCurrentAgent.

Fix

Introduce resolveSessionAgent(sess) — a single method that checks sess.AgentName first (set for background agent tasks) and falls back to CurrentAgent() for interactive sessions. This is used consistently across all affected call sites.

Additional fixes

• Nil-pointer panic: The previous inline resolution silently discarded the error from team.Agent() and could leave the agent nil. The new helper always falls back to CurrentAgent().
• Pre-loop event emissions: TeamInfo and ToolsetInfo events before the loop used r.CurrentAgentName() instead of the resolved agent.
• emitAgentWarnings / configureToolsetHandlers: These methods already received the agent as a parameter but independently called r.CurrentAgentName() for event names.
• Summarize: Used r.CurrentAgentName() and r.CurrentAgent() instead of the session-aware resolution.

Assisted-By: docker-agent

[docker-agent]

Review Summary

Assessment: 🟢 APPROVE

This PR correctly fixes a critical bug where background sub-agents were using the root agent's configuration instead of their own. The implementation is sound:

✅ resolveSessionAgent(sess) implementation is correct

• Properly checks sess.AgentName first for background agent tasks
• Safely handles errors from team.Agent() by falling back to CurrentAgent()
• Returns a non-nil agent in all cases

✅ All call sites correctly updated

• RunStream: Uses resolved agent consistently throughout the loop
• processToolCalls: Now uses session-pinned agent
• finalizeEventChannel: Correctly resolves agent for StreamStopped event
• Summarize: Properly resolves agent for compaction events
• configureToolsetHandlers & emitAgentWarnings: Now use the passed agent parameter instead of independently calling CurrentAgent()

✅ Eliminates race condition

• Background agent sessions no longer race on the shared currentAgent field
• Each session consistently uses its pinned agent throughout its lifecycle

✅ Nil-pointer safety improved

• Previous inline code silently discarded errors from team.Agent() which could theoretically leave agent nil
• New helper always returns a valid agent

Findings

No issues identified in the changed code. The refactoring is clean, the fix addresses the root cause, and the implementation handles edge cases correctly.